Page 57 - IJPS-7-2

International Journal of
Population Studies Child mortality by residence in Ethiopia

2.3. Statistical methods is based on multiplying regression coefficients using

We used STATA version 15.1 for statistical data processing regressors. And hence, to explain the urban-rural, intra-
and analysis. We checked missing values, and there were urban, and intra-rural inequalities in child mortality, we
no missing values for the outcome variable, but children used the Blinder-Oaxaca decomposition technique for a
with unknown birth places and children from not de jure non-linear variable. This technique allows for quantifying
the gap between the advantaged and the disadvantaged
residents were excluded from the analysis, which accounted groups (Ameyaw et al., 2021; Bazen et al., 2016; Fairlie,
for 3.2% of the cases. In this study, de jure residents refer to 2005; Jann, 2008; Bado & Appunni, 2015; Sinning et al.,
those residents who are usually living in a given area and 2008; Yaya et al., 2019).
who were counted as the resident of that area. We used
a correlation matrix to test multicollinearity effect of the 3. Results
explanatory variables using a cutoff of 0.6 known to cause
concern in multicollinearity (Senaviratna & Cooray, 2019). 3.1. Background characteristics of the study
Due to high and strong collinearity with a place of residence participants
(r = 0.8077) and sex of household head (r = 0.6236), the total Table 1 shows the background characteristics of the study
children ever born variable was removed from the model. population. Table 1 shows that 45.47% of children were
Place of residence cannot be removed since it is the key born to mothers residing in emerging regions and almost
identifier in investigating residential inequality in child the same proportion was born to mothers in developed
mortality. After removal of the total number of children ever regions, and 81.8% of children were born in rural areas.
born, an absolute correlation coefficient of <0.6 was observed The majority (78.56%) of respondent women were from
among predictors indicating the absence of multicollinearity. male-headed households; 83.09% of children were born
Although the variable ‘total children ever born’ was removed to households that had unimproved sanitation facilities;
from the model, the birth order was included in the model to only 5.28% of children were born to households that had
explore the relative effects of fertility on child mortality. clean cooking fuel; and 55.49% of households had six or

Considering the hierarchical nature of the 2016 EDHS more members. Table 1 also depicts that more than half
dataset into account, we used multilevel (i.e., three level: (54.8%) of children were born to households grouped
Community, household, and individual level) analysis with poor wealth status; 45.5% of children were born to
technique to get unbiased estimates of standard errors households that had improved sources of drinking water.
Most (70.7%) children were born at home; a slightly higher
and enable the modeling of between-level interactions by proportion (52.1%) of children were female. Table 1 also
treating every effect at the appropriate level. A multilevel illustrates that 19.7%, 43.9%, and 36.4% of children were
modeling explicitly accounts for the clustering of the units born in the 1 , the 2 – 4 birth order, and 5 and above
th
th
nd
st
of analysis. Besides, the multilevel modeling provides a birth order, respectively. More than 80% of mothers had
unified treatment for effects at individual, household, and initiated breastfeeding with their kids immediately after
community levels. Since the outcome variable is binary, birth. About 6% of mothers had given birth before entering
a multilevel logistic regression (Balluerka et al., 2010; the age of 18 years. More than 66% of children were born
Gelman & Hill, 2010) was used as a standard model for to uneducated mothers. Nearly 29% of children were born
assessing the effect of socioeconomic and contextual to mothers of Orthodox religion and 51.2% were born to
factors on child mortality in this study. Accordingly, four mothers of Muslim religion followers (Table 1).
models were fitted, including null model. The null model
was fitted to determine whether the use of multilevel 3.2. Results from multilevel analyses
modeling was appropriate in the analysis. Further, all
models were checked through interclass correlation Table 2 presents results from the multilevel regression
(ICC) and criteria information tests (AIC and BIC) and analysis. Here, it is crucial to understand that all individual
and household level factors are nested within the
their values were used to select the best model fitted for community (place of residence), hence, it needs to explain
multilevel analysis. We used likelihood ratio (LR) test to the residential inequalities using multilevel analysis. With
test statistically significant difference between two models this understanding, four multilevel models were fitted
based on the ratio of their likelihoods.
using only variables with P < 0.2 (Heinze & Dunkler, 2017)
We also used decomposition analysis to quantify the from the bivariate analysis (not presented in this paper).
contribution of observed and unobserved heterogeneity The overall multilevel analysis was conducted with random
at the individual, household, and community levels. intercept (only) model for both community and household
The decomposition analysis helps understand variance levels. First, the null model (Model 0, i.e., a model without
estimates whether regressors are random or fixed, which explanatory variables) was fitted and showed statistically

Volume 7 Issue 2 (2021) 51 https://doi.org/10.36922/ijps.v7i2.392